5. Developments

5.1 Equipment

• NVIDIA Pascal and Volta

  New generations of GPUs are seeing the widespread use of HBM (High Bandwidth Memory) stacked on the same substrate, reaching 32 GB and 1 TB/s throughput. This is also the case for FP16 and FP32 mixed-precision floating-point operators. Doubling performance at low 16-bit precision is particularly interesting for many applications in image processing and deep learning using neural networks, for example. AMD is also following this path.

• AMD Radeon 400 GPUs and APUs

  A major evolution concerns the integration of CPUs and GPUs on the same chip. After the Fusion range, AMD is now developing its APU (Accelerated Processing Unit) offering, with...